Offline Evaluation Testing

Overview

The @outputai/evals package provides an offline evaluation framework for testing workflow quality using datasets and evaluators. This is complementary to the runtime evaluator() from @outputai/core:

Aspect	Runtime Evaluators (@outputai/core)	Offline Eval Tests (@outputai/evals)
When	During workflow execution	After execution, at test time
Where	evaluators.ts in workflow folder	tests/evals/ in workflow folder
Purpose	Live quality scoring with confidence	Dataset-driven pass/fail verification
Triggered by	Workflow orchestration	output workflow test CLI command
Returns	EvaluationBooleanResult, etc.	Verdict helpers (pass/partial/fail)

Use offline eval testing when you want to validate workflow behavior against known datasets, build regression test suites, or assess subjective quality with LLM judges.

When to Use This Skill

• Creating files in tests/evals/ or tests/datasets/
• Writing evaluators that use verify() from @outputai/evals
• Creating YAML dataset files for test cases
• Building eval workflows with evalWorkflow()
• Running output workflow test commands
• Setting up ground truth data for evaluators

Directory Structure

Add a tests/ directory inside the workflow folder:

src/workflows/{workflow_name}/
├── workflow.ts
├── steps.ts
├── evaluators.ts          # Runtime evaluators (optional)
├── types.ts
└── tests/
    ├── datasets/
    │   ├── happy_path.yml
    │   └── edge_case.yml
    └── evals/
        ├── evaluators.ts  # Offline eval test evaluators
        ├── workflow.ts     # Eval workflow definition
        └── judge_topic@v1.prompt  # LLM judge prompts (optional)

Creating Evaluators with verify()

Import verify and Verdict from @outputai/evals (not @outputai/core):

// tests/evals/evaluators.ts
import { verify, Verdict } from '@outputai/evals';
import { z } from '@outputai/core';

verify() Signature

verify(options, checkFn)

Options:

• name — unique evaluator identifier (snake_case)
• input — Zod schema for the workflow input (optional, defaults to z.any())
• output — Zod schema for the workflow output (optional, defaults to z.any())

Check function receives:

{
  input,    // typed workflow input
  output,   // typed workflow output
  context: {
    ground_truth: Record<string, unknown>  // from dataset YAML
  }
}

Returns: any Verdict helper result.

Basic Example

import { verify, Verdict } from '@outputai/evals';
import { z } from '@outputai/core';

export const evaluateSum = verify(
  {
    name: 'evaluate_sum',
    input: z.object({ values: z.array(z.number()) }),
    output: z.object({ result: z.number() })
  },
  ({ input, output }) =>
    Verdict.equals(output.result, input.values.reduce((a, b) => a + b, 0))
);

Using Ground Truth

Ground truth values come from the dataset YAML and are available via context.ground_truth:

export const lengthCheck = verify(
  { name: 'length_check', input: blogInput, output: blogOutput },
  ({ output, context }) =>
    Verdict.gte(output.blog_post.length, Number(context.ground_truth.min_length ?? 100))
);

Verdict Helpers

All deterministic helpers return results with confidence 1.0.

Equality & Comparison

Method	Description
Verdict.equals(actual, expected)	Strict equality (===)
Verdict.closeTo(actual, expected, tolerance)	Within numeric tolerance
Verdict.gt(actual, threshold)	Greater than
Verdict.gte(actual, threshold)	Greater than or equal
Verdict.lt(actual, threshold)	Less than
Verdict.lte(actual, threshold)	Less than or equal
Verdict.inRange(actual, min, max)	Within inclusive range

String & Array

Method	Description
Verdict.contains(haystack, needle)	String includes substring
Verdict.matches(value, pattern)	Regex match
Verdict.includesAll(actual, expected)	Array contains all expected values
Verdict.includesAny(actual, expected)	Array contains at least one expected value

Boolean

Method	Description
Verdict.isTrue(value)	Value is true
Verdict.isFalse(value)	Value is false

Manual Verdicts

Method	Description
Verdict.pass(reasoning?)	Explicit pass
Verdict.partial(confidence, reasoning?, feedback?)	Partial pass with confidence
Verdict.fail(reasoning, feedback?)	Explicit fail

LLM Judge Evaluators

Before writing a judge prompt, identify the specific failure mode via error analysis (output-eval-error-analysis). Design the judge following output-eval-judge-prompt. After writing it, validate against human labels using output-eval-validate-judge.

For subjective quality assessments, use judge functions with .prompt files:

import { verify, judgeVerdict, judgeScore, judgeLabel } from '@outputai/evals';

// Returns pass/partial/fail verdict from an LLM
export const evaluateTopic = verify(
  { name: 'evaluate_topic', input: blogInput, output: blogOutput },
  async ({ input, output, context }) =>
    judgeVerdict({
      prompt: 'judge_topic@v1',
      variables: {
        blog_title: output.title,
        blog_post: output.blog_post,
        required_topic: String(context.ground_truth.required_topic ?? input.topic)
      }
    })
);

// Returns a numeric score from an LLM
export const evaluateQuality = verify(
  { name: 'evaluate_quality', input: blogInput, output: blogOutput },
  async ({ input, output }) =>
    judgeScore({
      prompt: 'judge_quality@v1',
      variables: { blog_title: output.title, blog_post: output.blog_post, topic: input.topic }
    })
);

// Returns a string label from an LLM
export const evaluateTone = verify(
  { name: 'evaluate_tone', input: blogInput, output: blogOutput },
  async ({ output }) =>
    judgeLabel({
      prompt: 'judge_tone@v1',
      variables: { blog_title: output.title, blog_post: output.blog_post }
    })
);

Judge .prompt File Format

Judge prompt files live alongside evaluators in tests/evals/:

# tests/evals/judge_topic@v1.prompt
---
provider: anthropic
# current as of 2026-05-04 — run output-dev-model-selection for the latest
model: claude-haiku-4-5-20251001
temperature: 0
maxTokens: 1000
---

<system>
You are an evaluation judge. Assess whether a blog post is faithfully about the required topic.

Return a JSON object with:
- verdict: "pass" if the blog clearly focuses on the topic, "partial" if it mentions the topic but lacks depth, "fail" if it is not about the topic
- reasoning: a brief explanation of your judgment
</system>

<user>
Required topic: {{ required_topic }}

Blog title: {{ blog_title }}

Blog post:
{{ blog_post }}

Judge whether this blog post is faithfully about the required topic.
</user>

Creating Eval Workflows

The eval workflow wires evaluators together and defines how to interpret results.

// tests/evals/workflow.ts
import { evalWorkflow } from '@outputai/evals';
import { evaluateSum } from './evaluators.js';

export default evalWorkflow({
  name: 'simple_eval',
  evals: [
    {
      evaluator: evaluateSum,
      criticality: 'required',
      interpret: { type: 'boolean' }
    }
  ]
});

Eval Definition Fields

Each entry in the evals array has:

• evaluator — the function created by verify()
• criticality — 'required' (affects pass/fail) or 'informational' (reported but doesn't block)
• interpret — how to convert the evaluator's return value into a verdict

Interpret Types

Type	Evaluator Returns	Mapping
{ type: 'boolean' }	Verdict.equals(), Verdict.gte(), etc.	true = pass, false = fail
{ type: 'verdict' }	judgeVerdict() or Verdict.pass/partial/fail()	Direct pass-through
{ type: 'number', pass: 0.7, partial: 0.4 }	judgeScore()	>=pass = pass, >=partial = partial, else fail
{ type: 'string', pass: ['a', 'b'], partial: ['c'] }	judgeLabel()	Label in pass list = pass, in partial list = partial, else fail

Full Example with Mixed Evaluators

export default evalWorkflow({
  name: 'blog_generator_eval',
  evals: [
    {
      evaluator: lengthOfOutput,
      criticality: 'required',
      interpret: { type: 'boolean' }
    },
    {
      evaluator: evaluateTopic,
      criticality: 'required',
      interpret: { type: 'verdict' }
    },
    {
      evaluator: evaluateQuality,
      criticality: 'required',
      interpret: { type: 'number', pass: 0.7, partial: 0.4 }
    },
    {
      evaluator: evaluateContent,
      criticality: 'informational',
      interpret: { type: 'boolean' }
    },
    {
      evaluator: evaluateTone,
      criticality: 'informational',
      interpret: { type: 'string', pass: ['professional', 'informative'], partial: ['casual'] }
    }
  ]
});

Naming Convention

The eval workflow name must end in _eval and match the pattern {workflow_name}_eval. The CLI resolves this automatically — output workflow test blog_generator looks for blog_generator_eval.

Dataset Files

For methodology on designing diverse datasets that cover failure-prone regions, see output-eval-dataset-design.

Datasets are YAML files in tests/datasets/. Each file represents one test case.

Basic Format

name: basic_input
input:
  values:
    - 1
    - 2
    - 3
    - 4
    - 5
last_output:
  output:
    result: 15
  executionTimeMs: 100
  date: '2026-02-13T00:00:00.000Z'

With Ground Truth

Ground truth provides expected values for evaluators. You can set global values and per-evaluator overrides:

name: stripe_blog
input:
  topic: "Stripe the payment processor"
  requirements: "Include a link to https://stripe.com/en-gb/pricing"
last_output:
  output:
    title: "Stripe: The Modern Payment Processing Platform"
    blog_post: |
      Stripe has revolutionized online payment processing...
  executionTimeMs: 5000
  date: '2026-02-16T00:00:00.000Z'
ground_truth:
  notes: "Known good case"
  evals:
    length_of_output:
      min_length: 100
    evaluate_topic:
      required_topic: "Stripe the payment processor"
    evaluate_content:
      required_content: "https://stripe.com/en-gb/pricing"

The ground_truth.evals.<evaluator_name> values are merged with the top-level ground truth and passed to the evaluator via context.ground_truth.

CLI Commands

output workflow test <workflow_name>

Runs evaluations against all datasets for a workflow.

Flag	Description
--cached	Use cached output from dataset files (skip workflow execution)
--save	Run workflow fresh and save output + eval results back to dataset files
--dataset <names>	Comma-separated list of dataset names to run (default: all)
--json	Output machine-readable JSON instead of the rendered report

Execution flow:

1. Loads all dataset YAML files from tests/datasets/
2. Without --cached: executes the workflow for each dataset to get fresh output
3. Sends all datasets to the {workflow_name}_eval workflow
4. Reports per-dataset and per-evaluator verdicts
5. Exits with code 1 if any required evaluator fails

output workflow dataset list <workflow_name>

Lists all datasets for a workflow with their cached status.

Flag	Description
--format <type>	Output format: table (default) or text
--json	Output machine-readable JSON

output workflow dataset generate <workflow_name> [scenario]

Generates a new dataset file by running the workflow.

Flag	Description
--input <json>	Workflow input as a JSON string or file path
--name <name>	Dataset filename (defaults to scenario name)
--trace <path>	Generate from a local trace file instead of running the workflow
--download	Download traces from S3 and convert to datasets
--limit <n>	Max traces to download from S3 (default: 5)

Common Usage

# Generate dataset from inline JSON input
output workflow dataset generate my_workflow --input '{"key": "value"}' --name my_test

# Generate from a scenario file
output workflow dataset generate my_workflow basic

# Run evals with cached output (fast, no re-execution)
output workflow test my_workflow --cached

# Run evals fresh and save results
output workflow test my_workflow --save

# Run specific datasets only
output workflow test my_workflow --dataset happy_path,edge_case

# List all datasets
output workflow dataset list my_workflow

Typical Workflow

# 1. Start the dev server
npm run output:dev

# 2. Generate datasets from real workflow runs
output workflow dataset generate blog_generator --input '{"topic": "AI"}' --name ai_post

# 3. Edit the dataset YAML to add ground_truth values for your evaluators

# 4. Run evals with --save to cache output and eval results
output workflow test blog_generator --save

# 5. Iterate on evaluators, re-run with cached output (fast)
output workflow test blog_generator --cached

# 6. List all datasets
output workflow dataset list blog_generator

Verification Checklist

• [ ] Evaluators import verify, Verdict from @outputai/evals (not @outputai/core)
• [ ] Eval workflow imports evalWorkflow from @outputai/evals
• [ ] All imports use .js extension
• [ ] Eval workflow name follows {workflow_name}_eval pattern
• [ ] Dataset YAML files are in tests/datasets/
• [ ] Evaluator files are in tests/evals/
• [ ] Each evaluator has a unique name in snake_case
• [ ] criticality is set to 'required' or 'informational' for each eval
• [ ] interpret type matches evaluator return type
• [ ] Ground truth keys in dataset match evaluator names
• [ ] Judge .prompt files are in tests/evals/ alongside evaluators
• [ ] z is imported from @outputai/core (not zod)

Related Skills

• output-dev-evaluator-function — Runtime evaluators using evaluator() from @outputai/core
• output-dev-scenario-file — Creating scenario JSON files for workflow execution
• output-dev-folder-structure — Understanding project directory layout
• output-dev-prompt-file — Creating .prompt files for LLM operations
• output-eval-error-analysis — Identify failure modes before building evaluators
• output-eval-judge-prompt — Design effective LLM judge prompts
• output-eval-dataset-design — Generate diverse test datasets
• output-eval-validate-judge — Validate LLM judges against human labels
• output-eval-audit — Audit an existing eval suite for trustworthiness